---
title: Anthropic's real-time safeguards
---

Canonical URL: https://ghost.fail/anthropic/real-time-safeguards

<TableOfContents />

# Anthropic's real-time safeguards

[Anthropic](/anthropic)'s infrastructure uses various safeguards to detect and mitigate risks in real-time. This page focuses primarily on classifiers.

## Classifiers

### Categories

- Cybersecurity
- CBRN (Chemical, Biological, Radiological, Nuclear)
- [User wellbeing](/anthropic/user-wellbeing)
- Underage users: "If a user under 18 self-identifies their age in a conversation, our classifiers will flag this for review and we'll disable accounts confirmed to belong to minors" Anthropic also reported actively developing a new classifier to detect subtle signs that a user might be underage. [^wellbeingpr]
- Frontier LLM research (see [Fable 5's safeguards](/models/claude-mythos-5/fable-safeguards))
- Copyright content filtering[^ipclassifier]

### Behaviors

#### Refusal responses

API requests that are blocked receive an HTTP 200 response with `stop_reason: "refusal"`.

#### System injections

See [the page on system reminders](/anthropic/system-reminders).

#### Fallback to less capable models

- [Claude Fable 5](/models/claude-mythos-5/fable-safeguards) -> Most recent Opus model ([Opus 4.8](/models/claude-opus-4-8) at launch)
- [Opus 4.8](/models/claude-opus-4-8) -> [Haiku 4.5](/models/claude-haiku-4-5)
- RSP-3 models ([Opus 4](/models/claude-opus-4), [Sonnet 4.5](/models/claude-sonnet-4-5)) -> [Sonnet 4](/models/claude-sonnet-4)

## Models' different safeguards

- [Opus 4.7](/models/claude-opus-4-7): (todo: document).
- [Opus 4.8](/models/claude-opus-4-8): Similar to Opus 4.7.
- [Fable 5](/models/claude-mythos-5/fable-safeguards): See page.
- [Sonnet 5](/models/claude-sonnet-5): Similar to Opus 4.8.

## Research

[^research1]: [Introducing Anthropic's Safeguards Research Team](https://alignment.anthropic.com/2025/introducing-safeguards-research-team/) (Alignment Science Blog, Feb 2, 2025)
[^research2]: [Constitutional Classifiers: Defending against universal jailbreaks](https://www.anthropic.com/research/constitutional-classifiers) (Anthropic, Feb 3, 2025)
[^research3]: [Monitoring computer use via hierarchal summarization](https://alignment.anthropic.com/2025/summarization-for-monitoring/) (Alignment Science Blog, Feb 27, 2025)
[^research4]: [Cost-Effective Constitutional Classifiers via Representation Re-use](https://alignment.anthropic.com/2025/cheap-monitors/) (Alignment Science Blog)
[^research5]: [Poisoning Fine-tuning Datasets of Constitutional Classifiers](https://alignment.anthropic.com/2026/backdooring-classifiers/) (Alignment Science Blog, Apr 24, 2026)
[^asl3safeguardsreport]: [AI Safety Level3 Deployment Safeguards Report](https://www-cdn.anthropic.com/dc4cb293c77da3ca5e3398bdeef75ee17b42b73f.pdf) (May 2025)
[^safeguardsdocs]: [Safeguards - Claude Support](https://support.claude.com/en/collections/4078535-safeguards) (Retrieved Jul 2, 2026)
[^cyberopussonnet]: [Real-time cyber safeguards on Claude Opus and Sonnet](https://support.claude.com/en/articles/14604842-real-time-cyber-safeguards-on-claude-opus-and-sonnet) (Retrieved Jul 2, 2026)
[^safeguardsnews]: [Building safeguards for Claude](https://www.anthropic.com/news/building-safeguards-for-claude) (Anthropic News, Aug 12, 2025)
[^wellbeingpr]: [Protecting the wellbeing of our users](https://www.anthropic.com/news/protecting-well-being-of-users) (Dec 18, 2025)
[^ipclassifier]: [Why am I receiving an 'Output blocked by content filtering policy' error?](https://support.claude.com/en/articles/9205721-why-am-i-receiving-an-output-blocked-by-content-filtering-policy-error) (Claude Support, Retrieved Jul 2, 2026)